Exploiting a Modified Gray Model in Back Propagation Neural Networks for Enhanced Forecasting

Abstract

Back propagation (BP) neural network has been widely used for various data predictions in application. One of the challenging issues in various applications of BP neural network is how to improve its reliability as well as its stability. In this paper, by using a prototype for the forecast of atmospheric radiation and atmospheric ozone concentration in the state of Ohio, USA, we show that a direct use of the BP neural network may lead to the loss of forecast reliability during its evolutionary process. Under the framework of Verhulst biological model with embedding cognitive computation, we show that one can effectively reduce the dispersion of pure randomness of data sets in BP neural network and improve the prediction of future data by incorporating accumulation generation operation (AGO) into the system. We demonstrate that the integration of a modified gray model and the AGO can offer a more desirable BP neural network algorithm, and both stability and reliability become much improved compared to the direct use of BP neural network reported in current literature.